Introduction

The recent COVID-19 pandemic resulted in more than seven million deaths worldwide1 and a meta-analysis of early 2020, including more than 280,000 individuals in 11 countries, reported a mortality about 6% of infected persons during the early stages of the pandemic, although underdiagnosis may overestimate the true infection fatality rate during this period2. However, solid-organ transplant (SOT) recipients were at higher risk of severe COVID-19 and death3,4,5,6,7,8 due to immunosuppressive therapies and frequent comorbidities, including cardiovascular disease, diabetes, and obesity6,9,10. This higher susceptibility to infection by other pathogens had been previously described and related to predisposing factors such as diabetes11, renal dysfunction12, surgery13, nutritional deficits14, or modifications of gut microbiota15 in addition to immunosuppressive treatment regimens16,17. Consequently, immune system alterations in SOT recipients, notably related to their immunosuppressive treatments17, are associated with lower vaccine responses18,19,20. As SOT recipients had impaired humoral and cellular immune responses to primary COVID-19 mRNA vaccination compared to healthy adults21,22,23, these patients were prioritized for booster doses of COVID-19 vaccine24,25.

However, despite vaccines being one of the fundamental tools in preventative medicine, the immunological correlates of effective vaccine responses remained largely unknown. Systems immunology has emerged as an indispensable tool to study the immune response in-depth in clinical settings. Combining multi-parameter data from high-throughput technologies and computational biology, systems immunology and vaccinology are able to unravel and predict the immune response under specific perturbation. Specifically, in systems vaccinology studies, key components of the cellular adaptive immune system, such as increased B cell activity with high plasmablasts within a week after vaccination, have been correlated with protective antibody levels in healthy individuals26,27. Importantly, this approach has highlighted the significant role of the innate immune compartment in vaccine response, which had been disregarded until then, where early type I interferon (IFN) response promotes protective antibody response after vaccination, specifically after receiving the influenza vaccine26,27. Interestingly, recent studies have integrated and compared the immune response of up to 13 different vaccines and, with the exception of few examples such as yellow fever, most of the vaccine responses elicit similar patterns with both higher adaptive immune response after one week and activation of the innate immune system already present at day 1 post-vaccination28,29. Similar studies have been conducted with COVID-19 vaccines, showing innate and adaptive immune responses after vaccination, the former notably enhanced after the second immunization as compared to primary vaccination with the BNT162b2 mRNA vaccine30,31. Further, using a systems vaccinology approach, CD38+ helper T cell subsets induced soon after the second vaccine dose were identified as a predictive marker of neutralizing antibody titers 1 month after the second vaccination31. These studies have been mainly performed in healthy adults, and the degree to which their results will hold true in particular at-risk patient subsets, such as SOT recipients, remains largely unknown.

Furthermore, the moderate efficiency of COVID-19 vaccination in SOT recipients and the paucity of data on the immune configuration at baseline in this immunocompromised population support the need for dedicated systems vaccinology studies. Here, we used a systems vaccinology approach, applying high-dimensional flow cytometry at baseline analyzed by computational modeling, to identify correlates of COVID-19 mRNA vaccine response in a cohort of 31 lung transplant recipients (LTR) and 59 kidney transplant recipients (KTR) recruited at the beginning of the COVID-19 vaccination campaign. Patients were vaccinated with 3 doses of Pfizer-BioNTech COVID-19 mRNA vaccine, and both logistic regression and machine learning approaches were used to identify demographic and clinical characteristics and baseline immune states correlated with humoral response to vaccination. Our results identified demographic and clinical features of the most at-risk patients for vaccine failure and highlighted the negative impact of mycophenolate mofetil (MMF) and more recent immunosuppression induction therapy on the humoral immune response to vaccination. We further identified a baseline immune state of SOT recipients characterized by an innate and activated B cell profile associated with an adequate vaccine response, whereas a T cell signature was associated with a poorer vaccine response after 3 doses of vaccine. We then extended our analyses to discriminate detectable and positive responders to 3 doses of vaccine. Finally, we estimated to what extent this baseline immune state was different from that associated with an early detectable humoral response after 2 doses of vaccination, highlighting the importance of double negative B cell subsets for this outcome.

Results

Systems vaccinology cohort design for identifying correlates of vaccine response in SOT recipients

To identify the pre-vaccination correlates for vaccine response in an immunosuppressed SOT cohort, 90 patients were recruited to a systems vaccinology study. The study population included 31 SARS-CoV-2 naïve LTR and 59 SARS-CoV-2 naïve KTR. Sex representation within these cohorts was 67.7% and 55.9% male, respectively, and the median age of participants was 59 and 63 years. The median time post-transplantation was 5 and 9.6 years in LTR and KTR, respectively. SOT recipients were on different combinations of immunosuppressive treatments, including corticosteroids (CS), antimetabolites such as azathioprine (AZA) and mycophenolate mofetil, calcineurin inhibitors such as tacrolimus (FK506) and cyclosporin A (CsA), and the mTOR inhibitor everolimus. Detailed demographic and clinical characteristics of the study cohort are shown in Table 1. Patients were vaccinated with the Pfizer-BioNTech COVID-19 mRNA vaccine and were categorized exhibiting no response (<5 binding antibody units (BAU)/mL of RBD (receptor binding domain) binding IgG), detectable response (detectable threshold at >5 BAU/mL) or positive response (positive threshold at >264 BAU/mL as used in previous studies32,33,34). Overall, vaccine response rates were low, as previously reported22,24. The SOT cohort had 0% of positive responses after 1 dose (assessed at day 21, before D2), increasing to 3.4% after 2 doses (assessed at day 49, 28 days after D2) and 24% after 3 doses (assessed at day 183 [182 - 195], 28 days after D3). Detectable responses were, however, observed in 6.6%, 39.8%, and 49.4% of the SOT cohort, after 1, 2, and 3 doses of vaccine, respectively (Supplementary Fig, 1). Of note, RBD binding IgG levels were not available in 2 KTR after 2 doses, 5 KTR after 3 doses, and 2 LTR after 3 doses. Four KTRs were excluded from the D3 analysis due to a documented SARS-CoV-2 infection occurring between D2 and D3. Additionally, 1 LTR experienced graft rejection during this period. Importantly, no significant changes in immunosuppressive treatments were made among SOT recipients during the study period.

Table 1 Demographics of study population
Full size table

This study design allowed for the determination of demographic, clinical, and baseline (pre-vaccination) immunological parameters that correlate with humoral immune responses, either detectable or positive, after 2 and 3 doses of the COVID-19 mRNA vaccine, respectively. Our study included 20 healthy controls to serve as a reference for baseline immune state, allowing comparison with SOT recipients. These healthy controls had not been tested for their serological response after vaccination, as the BNT162b2 vaccine had previously been associated with efficient seroconversion in virtually all healthy adults after 2 and 3 doses35,36. Their clinical characteristics are summarized in Table 1.

Specific demographic and clinical characteristics are associated with a deficient vaccine response in SOT recipients

We first investigated the effect of demographic and clinical variables on the vaccine response by logistic regressions (Table 2). No significant difference was observed between KTR and LTR when assessing detectable responses after 2 and 3 doses. However, a smaller proportion of LTR showed a positive response after 3 doses compared to KTR. Within all the SOT recipients, no significant effect was observed for BMI or ethnicity. Although not significant, a sex effect was observed, with males more likely to have a positive vaccine response after 3 doses. A complex relationship was observed with age. Higher vaccine responses were observed in patients younger than 50 years old, while vaccine responses were significantly reduced in patients aged between 50 and 70 years old, consistent with a known association of age with reduced vaccine efficacy37. This effect was, however, blunted in the oldest patient cohort, aged between 71 and 90 years old. We were also able to associate comorbidities of SOT recipients with vaccine response, with the only significant association being a detrimental effect of an estimated glomerular filtration rate (eGFR) below 30 mL/min/1.73 m2 after 2 doses. In addition, vaccine-induced immune response was impacted by immunosuppressive treatments, particularly in the case of a combination of multiple treatments for a positive response after 3 doses. However, as these drugs are ubiquitous within the poorly responding SOT cohort, effects could only be assessed in comparison to competing treatment regimens. Using this comparison, MMF treatment was the only significant influence over vaccine response, with significantly poorer outcomes after each dose, and using both the detectable and positive thresholds.

Table 2 Association between demographic and clinical variables with humoral response to COVID-19 mRNA vaccination in SOT recipients
Full size table

Baseline innate and activated B cell profile is associated with a positive response after 3 doses in SOT recipients, while a T cell signature correlates with a poorer response

Blood samples from each patient were analyzed by flow cytometry for 444 immunological variables, constituting a deep phenotypic analysis of the adaptive immune compartment with extensive innate immune profiling. We next investigated the association between baseline immunological profile (baseline immune state at day 0 before vaccination) and positive response to 3 doses of vaccine in SOT recipients. Multivariate logistic regression was performed to adjust for sex, age, type of transplant, and the number of years since transplantation. Of the immunological parameters associated with positive response to 3 doses after adjustment (Supplementary Data 1), the 10 parameters with greatest magnitude effects were further analyzed (Fig. 1A). The strongest positive associations were with non-classical monocytes and CD40+ plasmacytoid dendritic cells, within their respective parent populations, along with CD40+ median expression in CD21- B cells and double negative B cells DN2 (Fig. 1A) whereas the strongest negative associations were with T cells, half of them expressing HLA-DR. Then, we explored to what extent this baseline immune state was similar to that of healthy controls. For most of the baseline immune parameters, positive SOT responders after 3 doses had similar levels as healthy controls except for parameters on double negative CD4-CD8- T cells (Fig. 1B). These immune parameters were largely not directly impacted by MMF (Supplementary Fig. 2) or by CS treatment (Supplementary Fig. 3). To further explore this immunological configuration model, we performed two analyses of the global immune profile. First, we used a Uniform Manifold Approximation and Projection (UMAP) approach of the baseline immune state of SOT recipients, positive responders and non-responders after 3 doses, and healthy controls; considering the baseline immune profile of this latter group associated with a high humoral response to 3 doses of mRNA vaccination as reported previously36. Interestingly, the responders clustered closer to the healthy controls regarding all parameters (Fig. 2A) and the 10 most highly associated parameters after 3 doses (Fig. 2B), consistent with the individual parameter assessment described above. Second, we looked at the relationship between immune parameters using Spearman’s correlations. This allowed us to assess the relationship of the key immune parameters in healthy controls, and whether that relationship was preserved or disrupted in positive and non-positive responders SOT recipients. For the 10 most highly associated immune parameters with positive response after 3 doses, we observed strong correlations between different immune characteristics within the healthy controls, and these co-correlations were disrupted in a similar way among positive responders as compared to non-positive responders SOT recipients after 3 doses (Fig. 2C, D). Together, these analyses support a model where some SOT recipients exhibit a baseline immune profile closer to that of healthy controls in certain immune subsets, and allowing higher antibody response after 3 doses of vaccine.

Fig. 1: Identification of key immune parameters associated with positive humoral response to 3 doses of COVID-19 mRNA vaccination in SOT recipients.
figure 1

One hundred ten individuals were recruited to a systems vaccinology study, including 20 healthy controls, 31 LTR, and 59 KTR. Immunological profiles were assessed at baseline (day 0 of vaccination) for 444 immunological variables without any imputation. Multivariate logistic regression for each immunological variable was performed adjusting by sex, age, type of transplant and the number of years since transplantation, identifying the 10 highest magnitude parameters (based on log(OR of the z-score)) between significant (p-value < 0.05) associated with positive humoral response to 3 doses of COVID-19 mRNA vaccination. A Odds ratio (OR) of the z-score of the 10 highest difference parameters (magnitude cut-off), using the stratification of SOT recipients “positive” or “non-positive” responders according to anti-RBD binding IgG levels to 3 doses of COVID-19 mRNA vaccination. B Raw value graphs, showing healthy controls, dose 3 non-positive and dose 3 positive responders SOT recipients, for the 10 highest magnitude parameters associated with positive response to 3 doses of mRNA vaccination.

Full size image
Fig. 2: Baseline immune state and immunological relationships associated with positive humoral response to 3 doses of COVID-19 mRNA vaccination and machine learning approach predicting positive humoral response to 3 doses of COVID-19 mRNA vaccination in SOT recipients.
figure 2

One hundred ten individuals were recruited to a systems vaccinology study, including 20 healthy controls, 31 LTR, and 59 KTR. Immunological profiles were assessed at baseline (day 0 of vaccination) for 444 immunological variables without any imputation. Immune configurations displayed as UMAP plots for listed parameters for each individual, annotated based on patient characteristics. Distances between groups were calculated using the Calinski and Harabasz score and are shown on the right of each panel. For immunological relationships, multivariate logistic regression was performed adjusting by sex, age, type of transplant and the number of years since transplantation, identifying the 10 highest magnitude parameters (based on log(OR of the z-score)) between significant (p-value < 0.05) associated with positive humoral response to 3 doses of COVID-19 mRNA vaccination. Spearman’s correlations were calculated between each of the 10 highest magnitude parameters. For the machine learning approach, the support vector machine approach was selected to model vaccine response, based on the highest performance. A UMAP for all 444 all immune parameters, with individuals annotated as healthy controls, or dose 3 non-positive and dose 3 positive SOT recipients. B UMAP for 10 highest magnitude parameters associated with positive responses after 3 doses, with the same patient annotation. C For those immune parameters associated with positive response after 3 doses, Spearman’s correlations between each parameter pair are shown for healthy controls (left). For positive (center) and non-positive (right) after 3 doses in SOT recipients, the upper-right triangular matrix shows the Spearman’s correlations within the respective patient population, while the lower-left triangular matrix visualizes by color and size the deviations in the correlations observed between the SOT recipients and the healthy controls with green indicating high correlation within the healthy population and low correlation within the SOT population and brown low correlation within the healthy population and high correlation within the SOT population. D Values of the differential correlation for each cell population associated with a positive response to 3 doses, comparing the healthy controls and the respective SOT recipients as individual values (left) and violin plot (right). E ROC curve of the machine learning algorithm developed for predicting a positive response to 3 doses, as evaluated by nested cross-validation. Curve is mean performance in test groups during 10-fold cross-validation, with fill area the standard deviation. F Best features according to the machine learning model to identify the positive response to 3 doses of vaccine. The best features were chosen using the feature selection algorithm RFE-SVM, and the evaluation of the importance of each feature in the model was done using clustered permutation.

Full size image

B cells expressing activation markers are associated with positive responses to 3 doses of vaccine in SOT recipients

Our results show a baseline immune state associated with a positive response to 3 doses of vaccine in SOT recipients. To identify those immune parameters with the highest potential utility in diagnosis and to avoid redundancy, we used a machine learning approach. Multiple machine learning models were used and compared by recursive feature elimination, with 10-fold cross-validation to assess the explanatory power of the model, with the area under the Receiver Operating Characteristic (ROC) curve (AUC). Support vector machine (SVM) approach gave the highest accuracy (Table 3) with a predictive capacity (AUC) of 93.3% within the current cohort (Fig. 2E). To identify the immunological parameters contributing to the model, we used a cluster permutation approach. The two strongest single parameters were CD40+ median expression in memory B cell subsets, with CD40+ median expression in naive B cells also ranking highly, along with NK phenotypes such as CD38+ median expression in CD16-CD56dim NK cells (Fig. 2F). Notably, while there is some overlap with the logistic regression multivariate analysis used above, the lists of features are distinct, as highly co-correlated parameters provide decreasing amounts of additional information and therefore do not dominate the influential parameter list in machine learning approach. Together, this analysis suggests the key markers aiming to stratify SOT recipients into different functional groups regarding vaccine response to 3 doses of vaccine.

Table 3 Performance of different machine algorithms for COVID-19 mRNA vaccine responses in SOT recipients
Full size table

Comparison of baseline immune states and machine learning approach between detectable and positive response to 3 doses of vaccine in SOT recipients highlight differences regarding innate and adaptive immune cells

Beyond the ability to produce an efficient response to 3 doses of vaccine, we then examined the ability to reach our positive threshold (positive response to 3 doses) as compared to the ability to seroconvert without reaching the positive threshold (detectable non-positive response to 3 doses). Interestingly, activated and mature T cells were negatively associated whereas activated intermediate monocytes were positively associated with the ability to reach the positive threshold among detectable responders to 3 doses of vaccine (Supplementary Fig. 4A) (Supplementary Data 1). Again, baseline immune state and immune co-correlations of positive SOT responders after 3 doses were more similar to healthy controls as compared to detectable responders to 3 doses not reaching the positive threshold (Supplementary Fig. 4B, 5A–D). To discriminate SOT recipients able to reach the positive threshold among D3 detectable responders, CD40+ median expression in non-switched memory B cells was again the first discriminant parameter in our machine learning approach with 100% accuracy. Importantly, a T cell signature appears to be one of the most critical parameters distinguishing positivity status when compared to patients exhibiting detectable antibody levels following the third vaccine dose (Table 3) (Supplementary Fig. 5E, F).

Distinct baseline immune state dominated by double-negative B cell subsets is associated with the early detectable response after 2 doses in SOT recipients

After identifying a baseline immune state associated with a positive response after 3 doses, we investigated if the same basal immune state was already associated with an early vaccine response after 2 doses in SOT recipients. As very few SOT recipients had a positive response after 2 doses, we focused on the detectable response. Of the immunological parameters associated with detectable response to 2 doses after adjustment (Supplementary Data 1), the 10 parameters with greatest magnitude effects were further analyzed (Fig. 3A). The strongest positively associated parameter was CD80 expression, with CD80+ median expression on double negative CD27-IgD- B cells DN1 (CD11c-CD21+) and CD21- B cells including CD21-CD27+ B cells along with the frequency of CD80+ double negative B cells DN2 (CD11c+CD21-) (Fig. 3A). Surprisingly, for most of the baseline immune parameters, non-responders SOT had more similar levels to healthy controls than responders after 2 doses (Fig. 3B). Likewise, non-responders SOT clustered closer to the healthy controls regarding all parameters (Fig. 4A) and the 10 most highly associated parameters after 2 doses as compared to responders (Fig. 4B). We also observed more similar co-correlations between different immune characteristics at baseline between non-responders SOT and healthy controls than with responders (Fig. 4C, D). Finally, SVM was again the machine learning approach with highest accuracy (Table 3) with a predictive capacity (AUC) of 95% within the current dataset (Fig. 4E). The strongest single parameter was the frequency of CD40+ B cells and CD80+ median expression in switched memory B cells while the NK and NKT phenotypes such as CD38+ median expression in NKT and CD94+ median expression in NK and NKT cells had also an influence (Fig. 4F). Together, these analyses emphasize a distinct basal immune state associated with a detectable response to 2 doses as compared to a positive response to 3 doses of vaccine.

Fig. 3: Identification of key immune parameters associated with detectable humoral response to 2 doses of COVID-19 mRNA vaccination in SOT recipients.
figure 3

One hundred ten individuals were recruited to a systems vaccinology study, including 20 healthy controls, 31 LTR and 59 KTR. Immunological profiles were assessed at baseline (day 0 of vaccination) for 444 immunological variables without any imputation. Multivariate logistic regression for each immunological variable was performed adjusting by sex, age, type of transplant and the number of years since transplantation, identifying the 10 highest magnitude parameters (based on log (OR of the z-score)) between significant (p-value < 0.05) associated with detectable humoral response to 2 doses of COVID-19 mRNA vaccination. A Odds ratio (OR) of the z-score of the 10 highest difference parameters (magnitude cut-off), using the stratification of SOT recipients “detectable” or “undetectable” responders according to anti-RBD binding IgG levels to 2 doses of COVID-19 mRNA vaccination. B Raw value graphs, showing healthy controls, dose 2 undetectable and dose 2 detectable responders SOT recipients, for the 10 highest magnitude parameters associated with detectable response to 2 doses of mRNA vaccination.

Full size image
Fig. 4: Baseline immune state and immunological relationships associated with detectable humoral response to 2 doses of COVID-19 mRNA vaccination and machine learning approach predicting detectable humoral response to 2 doses of COVID-19 mRNA vaccination in SOT recipients.
figure 4

One hundred ten individuals were recruited to a systems vaccinology study, including 20 healthy controls, 31 LTR and 59 KTR. Immunological profiles were assessed at baseline (day 0 of vaccination) for 444 immunological variables without any imputation. Immune configurations displayed as UMAP plots for listed parameters for each individual, annotated based on patient characteristics. Distances between groups were calculated using the Calinski and Harabasz score and are shown on the right of each panel. For immunological relationships, multivariate logistic regression was performed adjusting by sex, age, type of transplant and the number of years since transplantation, identifying the 10 highest magnitude parameters (based on log (OR of the z-score) between significant (p-value < 0.05) associated with detectable humoral response to 2 doses of COVID-19 mRNA vaccination. Spearman’s correlations were calculated between each of the 10 highest magnitude parameters. For machine learning approach, the support vector machine approach was selected to model vaccine response, based on the highest performance. A UMAP for all 444 all immune parameters, with individuals annotated as healthy controls, or dose 2 undetectable and dose 2 detectable SOT recipients. B UMAP for 10 highest magnitude parameters associated with detectable responses after 2 doses, with the same patient annotation. C For those immune parameters associated with detectable response after 2 doses, Spearman’s correlations between each parameter pair are shown for healthy controls (left). For detectable (center) and undetectable (right) after 2 doses in SOT recipients, the upper-right triangular matrix shows the Spearman’s correlations within the respective patient population, while the lower-left triangular matrix visualizes by color and size the deviations in the correlations observed between the SOT recipients and the healthy controls with green indicating high correlation within the healthy population and low correlation within the SOT population and brown low correlation within the healthy population and high correlation within the SOT population. D Values of the differential correlation for each cell population associated with a detectable response to 2 doses, comparing the healthy controls and the respective SOT recipients as individual values (left) and violin plot (right). E ROC curve of the machine learning algorithm developed for predicting a detectable response to 2 doses, as evaluated by nested cross-validation. Curve is mean performance in test groups during 10-fold cross-validation, with the fill area the standard deviation. F Best features according to the machine learning model to identify the detectable response to 2 doses of vaccine. The best features were chosen using the feature selection algorithm RFE-SVM, and the evaluation of the importance of each feature in the model was done using clustered permutation.

Full size image

Discussion

Systems vaccinology studies have reshaped the understanding of vaccine function, by highlighting the role of innate immunity in vaccine protection. However, these studies have been conducted exclusively in healthy adults, with a notable lack of research applying this approach to vulnerable populations, including transplant recipients20. Our study is among the first efforts to help bridge this gap. Poor immune responses to COVID-19 mRNA vaccines have been widely documented in SOT recipients21,22,23,38. However, several studies have shown that additional boosters improve immune responses in immunocompromised populations39,40,41, underscoring the need for extra doses to enhance protection against SARS-CoV-2. Whether baseline immune configuration can predict the humoral response to COVID-19 mRNA vaccine among SOT recipients has remained largely unexplored. In our study, 24% of SOT recipients were able to induce an effective response after 3 doses of mRNA COVID-19 vaccine. One of the goals of our study was to identify an immune signature capable of predicting vaccine responders within our current cohort. We specifically found a pattern which showed a baseline immune state dominated by innate and B cells expressing activation markers, whereas a T cell signature was associated with a poorer vaccine response. The low rates of seroconversion we observed in our SOT cohort after mRNA vaccination are consistent with previous studies showing KTR and LTR the poorer responders to COVID-19 vaccination among SOT recipients42. Although KTR were more likely to mount a vaccine response above the positive threshold after the third dose in our study, we grouped together both SOT cohorts in order to achieve patient numbers sufficient for downstream statistical analysis. Beyond the low humoral response induced by vaccination, we wanted to investigate the impact of demographic and clinical variables on vaccine response. Men showed a higher positive vaccine response after dose 3, although not significantly, as Masset et al. showed among 136 KTR43, whereas antibody responses to bacterial and viral vaccines have been shown to be generally higher in women44. Consistent with a known association of age with reduced COVID-19 vaccine-induced immune responses in the general population37,45 and in SOT populations46, SOT recipients aged between 50 and 70 years old showed lower immune responses than SOT recipients younger than 50 years old. However, this association was dampened in SOT recipients aged over 70 years old. Although older SOT recipients showed a higher proportion of old grafts (above 10 years post-transplant) (Supplementary Fig. 6A) and took a lighter immunosuppressive regimen (Supplementary Fig. 6B), both potentially associated with a better immune response to vaccination46,47,48 given the profound immunosuppression caused by induction treatment49,50 and multiple chronic immunosuppressive treatments, SOT recipients older than 70 years old still showed a better immune response than between 50 and 70 years old after adjustment for these parameters. This could be related to a selection bias for survivors of infections, suggesting a more efficient residual immune response in this older group. Moreover, an eGFR below 30 mL/min/1.73m2 was negatively correlated with the vaccine response, consistent with previous reports48. Indeed, uremia has been shown to be associated with a state of immune dysfunction caused on the one hand by the accumulation of pro-inflammatory cytokines due to decreased renal elimination and increased production as a result of uremic toxins, oxidative stress, volume overload, and comorbidities51,52 and, on the other hand, by the impact of the uremic milieu and a variety of associated disorders exerted on immunocompetent cells53. According to previous publications42,48,54,55, a high number of immunosuppressive treatments and MMF were associated with a lower vaccine-induced immune response. MMF inhibits purine synthesis by reversible inhibition of inosine monophosphate dehydrogenase, in particular the type II enzyme, which is expressed predominantly in activated lymphocytes56 and has become widely used in SOT populations to prevent allograft rejection in addition to other immunosuppressive treatments, improving the efficacy profiles of immunodepression counterbalanced with detrimental toxicities57,58. Given this profound immunosuppression and the subsequent lower vaccine response in SOT recipients, alternative strategies such as a temporary discontinuation of MMF have been explored in SOT cohorts to increase the immunogenicity of COVID-19 vaccines, but revealed controversial results59,60,61.

Regarding specifically the immune cell subsets involved in predictive vaccine response, a few interesting observations can be reported. Focusing on B cells, while total B cell numbers are generally decreased in SOT recipients62,63, it is interesting to note that most of the baseline parameters associated with antibody response either after 2 or 3 doses rely on activation markers (CD80 and CD40) on double negative (DN, i.e. CD27-IgD-) B cell subsets. DN B cells are mainly increased in the context of aging, infectious or autoimmune diseases64, where they have been directly implicated in auto-antibodies secretion in systemic lupus erythematosus65. More importantly, these “unconventional B cells” have been involved in influenza vaccine response66 and their expansion has been directly linked to protective immunity after influenza vaccine in a younger cohort in comparison to an elderly cohort67. While increased DN B cells have been reported in other SOT cohorts63, we mainly observed an increase in DN B cells at baseline in patients responsive after both 2 and 3 doses of COVID-19 vaccine in our cohort (Supplementary Fig. 7A). Moreover, activation markers such as CD86 have been reported to be decreased in SOT recipients62,63 and it is therefore interesting to notice the best vaccine response in patients expressing a higher level of activation markers (CD80 or CD40) on their B cells. Interestingly, total transitional B cells (CD19+CD27-CD38++CD24++) were severely decreased in all our SOT recipients in comparison to healthy controls (Supplementary Fig. 7B) as reported earlier62,63. Previous studies have suggested a correlation between baseline levels of transitional B cells and humoral response to COVID-19 vaccine in KTR63,68. In our study, we could not confirm transitional B cells as a predictive parameter, probably due to the very low levels in SOT recipients. A correlation of responders SOT recipients and slightly higher levels of transitional B cells was observed, but was not significant in our cohort (Supplementary Fig. 7B).

Baseline activation of the innate immune cells has been linked to protective antibody response after vaccination in various studies29,69. The significance of this phenomenon stems from the necessity of robust innate immune activation to effectively prime the adaptive immune response. Therefore, there is a high probability that our observed association with elevated CD40+ pDC and activated monocytes in our SOT cohort contributes to the effective antibody production after vaccination against COVID-19. The individual case of monocyte activation might require specific attention. In particular, a pro-inflammatory signature is associated with higher antibody levels in young healthy adults, while the same inflammatory profile has been associated with a worse response in older adults29,69. In our cohort, non-classical monocytes were associated with a positive antibody response after the third vaccine dose. As non-classical monocytes are mainly described as anti-inflammatory cells70, there is a possibility that our SOT recipients behave as older adults with a pro-inflammatory myeloid signature being detrimental to antibody response. However, activated intermediate monocytes, known to be rather pro-inflammatory cells, were also important to differentiate the positive from the detectable responders after the third dose of vaccine. Therefore, further studies are needed to clarify the impact of monocyte activation status in the vaccine response of SOT patients. Importantly, a recent meta-analysis combining transcriptomic data from 13 different vaccines from 28 studies has identified a common baseline immune profile predicting protective antibody levels across various vaccines in healthy controls29. There, the baseline immune signature represented by higher abundance of adaptive cell subsets (as opposed to innate pro-inflammatory markers) was generally associated with a lower B cell response on day 7 and overall lower antibody production after vaccination. It is possible that this signature partially overlaps in SOT patients, as our study found that SOT responders after the third exhibited an immune signature more similar to that of healthy controls. Interestingly, in our dataset, we observed as well that most of the markers related to T cells were associated with low response after the third vaccine dose. These markers were related to different T cell subsets, mainly activated cells (HLA-DR+) or with a more mature (CD57+) phenotype.

Using a machine learning approach to investigate our results, there were several baseline parameters within the NK cell subsets that could contribute to the prediction of vaccine responders in our cohort, after both the second and third dose. Little is known about the role of NK cells during a vaccine response. They could contribute to the early immune system activation with their innate function through cytokine secretion, as observed during the response after vaccination against Ebola71. Additional roles with adaptive memory function have been attributed to NK cells during specific responses after vaccination against HIV or HBV72,73. Specific subsets, such as CD56dimCD57+ cytotoxic NK cells, are mediating this response and similar subsets (CD56dimNKG2C+) at baseline have been recently identified to predict antibody response after COVID-19 vaccine in healthy adults74. In our cohort, higher baseline activation makers such as CD57, CD94, or CD38 seem to be predictive for a higher antibody level. However, further focused investigation of the role of NK cells in vaccination is required as NK cells can play a double role with regards to their suppression of activated CD8+ and CD4+ T cells through their cytotoxic activity during vaccine immune responses, as they negatively correlate with protection after adjuvanted vaccine against HBV or malaria75,76.

While this study identifies different immune correlates associated with efficient response to 3 doses of mRNA COVID-19 vaccination in 90 SOT recipients, allowing for in-depth statistical analysis, there are several important limitations to note, on the cohort design, study design, and statistical power. First, on the cohort design, we have a limited sample size, due to the monocentric recruitment of SOT recipients. We note, however, that the poor vaccine response in SOT recipients previously reported in other studies could be confirmed in our study, indicating our cohort is representative. Additionally, as the immune landscape in SOT recipients is highly influenced by immunosuppressive treatments received during and after transplantation, and is selected based on patient-specific criteria, our cohort presents as a heterogeneous group of patients. The frequent simultaneous administration of different treatments makes it difficult to dissect the effect of single treatments. Further, sampling restrictions in healthy individuals precluded follow-up serology, and it is thus possible that a small number of non-responders may have been present within the healthy control population. While response after two doses is near-ubiquitous in healthy individuals35,36, without direct measurement, we cannot exclude demographic differences resulting in a minor non-response rate. Second, on study design, the rationale selection of biological features to be measured in this study resulted in the exclusion of several potentially interesting parameters. In particular, antigen-specific T cell responses were not measured, due to the incompatibility of cohort powering with HLA-typing. As some SOT recipients classified as non-responders due to absent seroconversion may, in fact, have generated antigen-specific memory T cell responses, the relationships identified here can be considered associations with seroconversion rather than vaccine-response. Humoral virus-specific antibody responses are currently, however, the best validated correlate of protection against SARS-CoV-2 infection in the general population and likely in SOT recipients as well77. It is notable, however, that several studies have reported a dissociation between B and T cell responses in SOT recipients following COVID-19 vaccination78,79,80,81. While virus-specific CD8+ T cell responses remain generally low in SOT recipients78,79, the expansion of virus-specific CD4+ T cells has been observed even in the absence of a robust anti-RBD antibody response78. Future studies should therefore incorporate the evaluation of cellular immune responses as additional outcomes, particularly in the context of systems vaccinology approaches in SOT cohorts. Finally, on statistical power, due to the intrinsic design limitations of systems immunology studies and the methods employed here, multiple testing could not be taken into account, and there remains a possibility of statistical overfitting. While we have implemented rigorous internal validation strategies, our predictions have yet to be confirmed in an independent cohort. We acknowledge the importance of external validation and consider it a critical next step to further substantiate our findings. Therefore, additional systems vaccinology studies in other SOT populations are needed to validate our observations.

An important outcome from this study was the identification of functional immune configurations present in SOT recipients. An immune configuration can, in this case, be defined as a shared meta-phenotype, where individuals share multiple small modifications in immune parameters, which translate into potentially large differences in functional outcomes82. SOT recipients have undergone multiple severe immunological challenges; the potential initiating pathology, the transplantation with induction treatment, the acute and chronic immune reactions against the graft, and the continued use of immunosuppressive drugs. With the poor response to vaccination exhibited in these patients, a simple model may postulate a single gradient of immune profile, ranging from dysfunctional through to the fully recovered and functional immune state shared with healthy individuals. This model can be supported, to some extent, by the data shown after 3 doses of vaccine. We identified patients with an effective response and a basal immune configuration closer to that of healthy controls as compared to non-responder patients. However, we identified a distinct immune configuration that is shared by some SOT recipients capable of an early detectable immune response after 2 doses. This immune configuration does not fit within a simple linear immune configuration model, instead achieving a baseline immune state more deviant from that of healthy controls. It is tempting to speculate that SOT recipients exhibit complex immune configurations with a baseline immune state dominated by B cells, including DN B cells, associated with an early detectable humoral response to 2 doses, and a different baseline immune state with low T cell activation imprint associated with a late efficient humoral response to 3 doses of vaccine. While our study focuses on the outcome of the primary response to the COVD-19 mRNA vaccine, it remains possible that our predictions, after external cohort validation, could also be applied to non-responders in a post-pandemic setting, where the major immune response is held by secondary response due to repetitive exposure to SARS-CoV-2 and vaccination boosters. Furthermore, as universal baseline immune signatures predicting vaccine responses across different vaccines have been identified in healthy individuals29, our findings may be extrapolated to predict immune responses to various vaccines in the vulnerable SOT population. In addition, our approach could serve as a strategy for patient stratification at baseline to predict vaccine response, predicting individuals who would remain susceptible after the standard vaccine regimen and who would benefit the most from additional booster doses. Further systems vaccinology studies in SOT recipients would be highly informative to verify these observations. Not only could they provide potential diagnostic value for the patients, but knowledge on how these immune configurations are shaped and their stability over time may provide broader understanding of how optimal immune responses to vaccination are achieved.

Methods

Study design and participant characteristics

The study was designed to assess demographic, clinical, and baseline immunological parameters predicting the humoral response to BioNTech/Pfizer BNT162b2 mRNA (Comirnaty®) vaccination in LTR and KTR, naive for SARS-CoV-2 infection. A cohort of health-care workers from Hôpital Saint-Pierre in Brussels was used as a control group regarding baseline immune parameters.

The ethics committee of Hôpital Erasme, Brussels, Belgium (references P2021/182 and A2021/156 for LTR; P2020/284 and A2021/131 for KTR) and the Belgian Federal Agency for Medicines and Health Products (FAMHP, EudraCT 2021-004992-16 for LTR; EudraCT 2021-000-412-28 for KTR) approved the monocentric prospective phase IV investigator-initiated study of the immunogenicity of the BNT161b2 vaccine (Pfizer-BioNTech). The ethics committee of Hôpital Saint-Pierre approved a 6-month study on the carriage and seroprevalence of SARS-CoV-2 in health-care workers (CE/20-04-17)83.

All participants of at least 18 years of age provided written informed consent for each study, respectively. LTR and KTR from the Chest department and Nephrology, dialysis, and transplantation department in Hôpital Erasme, Belgium, respectively and health-care workers from Hôpital Saint-Pierre were enrolled before COVID-19 vaccine administration. Participants with a previous laboratory-confirmed SARS-CoV-2 infection or with a baseline anti-RBD IgG level >5 BAU/mL were considered previously infected and excluded from this study. Demographics of the healthy and SOT populations are given in Table 1. SOT recipients were vaccinated with the Pfizer-BioNTech COVID-19 Vaccine on the day of blood sampling (baseline), with second and third doses given on days 21 and 155 [154 – 167], respectively. SOT recipients and health-care workers had blood drawn to collect peripheral blood mononuclear cells (PBMCs) just prior to the first vaccine dose (baseline). SOT recipients had blood drawn to collect serum three weeks after D1 prior to the second dose (post-D1), four weeks after the second dose (post-D2), and four weeks after the third dose (post-D3). Health-care workers were not sampled after vaccination and are considered in our study as the phenotype reference, according to the very high rates of seroconversion of healthy adults after mRNA vaccination, as reported previously35,36.

Sample processing and isolation of PBMCs

Serum tubes were centrifuged at 3000 rpm for 10 min to separate serum. Serum was stored at -80°C for antibody analysis. PBMCs were isolated from heparinized whole blood by density gradient centrifugation as recommended (Leucosep tubes, Greiner). Whole blood was diluted with an equal amount of Dulbecco’s Phosphate-Buffered Saline (PBS) with 2% Fetal Bovine Serum (FBS) and centrifuged. The upper plasma layer was discarded without disturbing the interface. The PBMCs layer was retained and washed three times with medium. Cell pellets were diluted with HBSS and aliquoted. PBMCs were frozen progressively with 10% Dimethyl Sulfoxide (DMSO) in FBS and stored in liquid nitrogen.

SARS-CoV-2-specific binding antibodies by enzyme-linked immunosorbent assay

Binding antibodies at baseline and after vaccination were assessed manually using an enzyme-linked immunosorbent assay (ELISA) for the quantitative detection of IgG-class antibodies to RBD (Receptor Binding Domain, Wuhan strain) [Wantai SARS-CoV-2 IgG ELISA (Quantitative); CE-marked; WS-1396; Beijing Wantai Biological Pharmacy Enterprise Co., Ltd, China] as previously described84,85. For quantification of antibodies, diluted serum samples (1/10, 1/100, 1/400, 1/1600 and 1/6400) were tested with an internal standard, calibrated against NIBSC 20/136 (First WHO International Standard Anti-SARS-CoV-2 Immunoglobulin), and an external positive control sample was included on each plate. Diluted samples were incubated (37 °C, 30 min) with pre-coated microwells and washed five times. Next, plates were incubated with horseradish peroxidase (HRP)-conjugated anti-human IgG antibodies (37 °C, 30 min) and washed five times before adding a TMB and urea peroxide solution for 15 min (37 °C, dark). After incubation, a stop solution (0.5 M H2SO4) was added, and optical density (OD) was measured at 450 nm using a Bio-Rad iMark Microplate Absorbance reader (cat#1681130). Net OD values were converted to arbitrary IgG units per mL by interpolation from a point-by-point plot fitted with the standard concentrations and net OD values (correlation coefficient R2 ≥ 0.9801), using GraphPad Prism version 9.0.0 for Windows (GraphPad Software, San Diego, California, USA) and exported to Microsoft Excel. Antibody measurements were adjusted for any sample dilution, converted to binding antibody units per mL (BAU/mL) and reported as such. The lower limit of quantification was 5 BAU/mL according to the WHO standard. Clinical performance characteristics of the assay, evaluated in 69 PCR-confirmed COVID-19 patients (comprising mild and severe clinical outcomes, ≥ 15 days post onset of symptoms) and 167 pre-pandemic sera, resulted in a specificity of 100% (95% CI 97,75-100) at a sensitivity of 100% (95% CI 94,73-100) for a cut-off of 6 BAU/mL. Data lower than the limit of detection were attributed the value 2.5 BAU/mL before transformation. Responses reaching the threshold of >264 BAU/mL, as previously used32,34, were considered “positive”, while responses >5 BAU/mL were “detectable”.

Flow cytometry

Frozen PBMCs were thawed in complete RPMI, washed twice with PBS, and stained with live/dead marker (fixable viability dye eFluor780; eBioscience) as well as fluorochrome-conjugated antibodies against surface markers: anti-CD3 (REA613, Miltenyi Biotec), anti-CD123 (7G3), anti-CD80 (L307.4), anti-CD21 (B-ly4), anti-CD27 (L128), anti-BAFF-R (11C1), anti-CD94 (HP-3D9), anti-CD8 (SK1), anti-CD86 (2331 (FUN-1)), anti-CD141 (1A4), anti-CD56 (NCAM16.2), anti-CD4 (SK3), anti-CD16 (3G8), anti-CD40 (5C3, all BD Biosciences), anti-CD57 (HNK-1), anti-CD24 (ML5), anti-HLA-DR (L243), anti-CD19 (HIB19), anti-IgM (MHM-88), anti-CD11c (3.9), anti-IgD (IA6-2), anti-CD10 (HI10a), anti-CD38 (HB-7, all BioLegend), anti-CD14 (TuK4, eBioscience). Cells were then fixed using the Foxp3/Transcription factor buffer staining set (eBioscience), and data were collected on a BD FACSymphony A5 Cell Analyzer (BD Biosciences). A maximum of 5×105 events were acquired for each sample. The data of 5 KTR were excluded from downstream analysis due to insufficient sample quality, including low cell viability.

A subset of samples was also used to perform an additional staining using markers focused on the T cell compartment. Cells were thawed and plated as described above, before staining with a live/dead marker (fixable viability dye eFluor780, eBioscience) was performed. After washing, cells were stained with fluorochrome-conjugated antibodies against surface markers: CD14 (61D3), CD4 (RPA-T4), CD45RA (HI100), CCR7 (3D12), CXCR5 (MU5UBEE), ICOS (ISA-3, all eBioscience), CD8 (RPA-T8, Biolegend). Cells were washed again and fixed using the Foxp3/Transcription factor buffer staining set (eBioscience) before acquisition on a BD Canto II (BD Biosciences). A maximum of 2.5 × 105 events were acquired for each sample. Raw.fcs data were compensated using AutoSpill86. The complete set of FCS files used has been deposited on FlowRepository and annotated in accordance with the MIFlowCyt standard. These files may be downloaded for further analysis from http://flowrepository.org/id/FR-FCM-Z662.

Flow cytometry gating

Compensated .fcs files were manually pre-processed in FlowJo (LLC, Ashland, Oregon; v10.8) to only contain events with stable fluorescent signals. For samples acquired on the FACSymphony, gating on singlets, leukocytes, live cells, and subpopulations was performed in R (version 4.0.2) by concatenating 7 individual samples and automatically calculating adequate boundaries between populations using the gating strategy outlined in Supplementary Data 2 and Supplementary Fig. 8. These boundaries were then applied to every individual sample to determine cell frequencies and median fluorescence intensity (MFI) values of activation markers on relevant populations. A minimum of 50 cells was necessary to define a cell population. (Neumann et al., manuscript in preparation). For samples acquired on the BD Canto II, manual gating was performed selecting singlets, lymphocytes, and living cells before excluding CD14+ monocytes. CD4+ and CD8+ T cells were further divided based on their CCR7/CD45RA expression profile, and expression of ICOS and CXCR5 was evaluated on non-naive cells (excluding CCR7+CD45RA+ cells). For the gating strategy, see Supplementary Fig. 9.

Association between demographic and clinical variables with vaccine response

To assess the association of demographic and clinical characteristics with the COVID-19 vaccine response, sex and age were considered a priori confounders, i.e., always considered confounders regardless of their effect on the relationship. The other demographic variables and treatment regimens were considered potential confounders. They were only considered confounding if a significant impact on the model between the exposure evaluated and the outcome was observed. The relationship was evaluated using multivariate logistic regression. Stepwise regression bidirectional elimination was used to select among potential confounders. The stepwise criterion of permanence was a significant p-value (<0.05) of the potential confounder, thus characterizing its importance for the model.

Batch correction

For flow cytometry, a batch variation in channel intensity is inherent to the measurement process. Here, we first verified whether there were batch differences in the cytometry data. We found flow cytometry readouts with a significant difference between batches (t Student/Mann-Whitney test), while other demographic and clinical variables of the study did not show significant differences between batches (t Student/Mann-Whitney test for continuous variables and Pearson’s chi-squared for categorical variables). This indicated that the batch effect differences were due to technical variation in the flow cytometer and staining procedure. To correct measurement distortions due to day-to-day variation, we located a cell population that had the smallest proportional variation between samples for each immune marker (considering populations of cells with more than 30 cells) and used their mean to correct the respective marker channel.

Identification of immune variables associated with vaccine response

For each of the 444 immunological variables, we assessed their association with vaccine response by considering only SOT recipients divided based on: (i) the presence of antibodies against SARS-CoV-2 above the positive threshold/below the positive threshold after the third dose among all SOT recipients; (ii) the presence of antibodies against SARS-CoV-2 above the positive threshold/below the positive threshold after the third dose among detectable SOT recipients; (iii) the presence of antibodies against SARS-CoV-2 detectable/undetectable after the second dose among all SOT recipients. This stratification of patients was based on the very low number of SOT recipients reaching the positive threshold, particularly after only two doses, requiring analysis to focus on achieving the detectable antibody reading, or on reaching the positive threshold after the third dose only. As each immune marker may have different levels of measurements, to compare the magnitude of the effect of each marker on the vaccine response, a z-score transformation was applied. The evaluations of the associations of the immune markers in vaccine response were performed through multivariate logistic regression. Sex and age were considered a priori confounders. Stepwise regression bidirectional selection was used to select additional potential confounders. The stepwise criterion for selecting features was Akaike information criterion87, keeping only potential confounders that improve the model. This resulted in adjustments for sex, age, type of transplant, and the time since transplant. Among the immunological markers that obtained p-values lower than 0.05, the 10 that had the greatest magnitudes of effect were chosen for further analysis. Multiple testing correction was not applied, as our analysis focused on the interpretation of a group of markers, with each marker considered independently. While this approach increases the likelihood of identifying biologically-relevant features in a study design where the number of features outweighs the size of the cohort, the approach will identify false positives. For evaluation of the confidence intervals associated with each feature, see Supplementary Data 3 and 4. Logistic regression models were run in Python using the package statsmodels version 0.13.2.

Similarities between immune markers in SOT recipients and healthy controls

To understand the similarity of the immune response between SOT recipients and healthy controls, the UMAP dimension reduction technique88 was performed. To do so, markers with more than 30% of missing values were first removed, and subsequently, the remaining missing values were imputed using the mean. After that, a standardization transformation was performed. To assess the distances between individuals, the Calinski and Harabasz score89 was used. In addition to understanding the immune similarity between SOT recipients and healthy controls, it is also important to understand whether the relationships between immune markers change within individuals. For that, we used a heatmap for Spearman correlation and compared these correlations in SOT recipients with healthy controls.

Machine learning

Data was cleaned and standardized as described above, with standardization applied within each group. For selection of important markers contributing to the outcome by machine learning, we used recursive feature elimination90, with 10-fold cross-validation using the area under the Receiver Operating Characteristic (ROC) curve as readout. The model used for feature selection was a support vector machine (SVM). The choice of model hyperparameters was performed by grid search with 3-fold cross-validation using balanced accuracy as a selection criterion.

Due to the multicollinearity shared between the markers, the potential importance of the marker can be underestimated, as there are other similar markers contributing similar information. Therefore, to limit this potential loss of important markers, we applied hierarchical clustering of markers based on the dissociation of the markers given 1 minus the modulus of the Spearman correlation. After the identification of these clusters, the assessment of the importance of each marker was performed using permutation importance91 using all markers and excluding markers contained in the same group as the evaluated marker. Thus, the marker’s importance in the phenotype is its impact on the model containing only itself in its group and a marker from every other group. The model used in the permutation presented the same hyperparameters selected in the feature selection. Permutation was performed with 100 repetitions and evaluated by the area under the ROC curve (AUC).

The machine learning models were created using the best features to reduce dimensionality. The best models were selected and evaluated by nested cross-validation, applying 10 external and 5 internal cross-validations. The selection of the hyperparameters of the models was done by grid search using balanced accuracy as a criterion.